The rapid growth in satellite imagery has helped scientists understand the Earth better. The improved understanding of the Earth makes it possible for scientists to perform better in all activities that range from disaster management in the form of mobilizing resources to comprehend global warming by monitoring its effects. The major limitation of this achievement is the assumption that significant features in satellite images, like buildings, roads, trees, or water bodies, can be easily identified, either manually or semi-automatically, but always perfectly. In this paper to overcome this limitation, we use different convolutional neural networks with modifications such as proposed PSPNet, U-net architecture, Inverted pyramid and XGBoost algorithm for accurately detecting specified features in satellite images from Defense Science and Technology Laboratory (DSTL) database. Automation of feature detection in satellite images is not only useful in making smart and quick decisions, but also in bringing innovation in the application of computer vision methodologies to satellite imagery.